4. Agentic AI: What Changes When Software Acts

Most AI you have encountered is reactive. You give it an input, it returns an output, and it waits. A chat interface is the obvious example. It does nothing until spoken to, and it does nothing after answering.

An agentic system is one that pursues a goal. Give it an objective, and it breaks the objective into steps, decides which tools or systems to use for each step, carries them out, checks the result, and adjusts. It does not wait for a prompt at each stage. This has become practical fairly recently, driven by models that can reason through a sequence of actions rather than producing a single answer.

The distinction sounds technical. It is actually the most consequential thing in this series.

The step that changes everything

A reactive system informs. A person reads the output and decides what to do.

An agentic system acts. It drafts the notice and sends it. It reads the specification and updates the record. It receives the enquiry, asks the customer for what is missing, prepares the quotation, and puts it in front of a reviewer. The human is no longer the mechanism by which anything happens; the human is now a supervisor of something that happens whether or not they are paying attention.

Every serious question about enterprise AI follows from that shift. Not from the intelligence of the system, and not from its accuracy, but from the plain fact that it is now doing things.

The hard part is not autonomy

The industry conversation treats autonomy as the achievement. More autonomous is understood as more advanced, and vendors compete on how much the system can do without a human.

Our experience points the other way. The difficult part of an agentic deployment is not making the system capable of acting. That is largely a solved engineering problem. The difficult part is deciding, precisely, what it is allowed to do, and building those limits into the architecture so they hold when nobody is watching.

We call this the mandate, and we have written about it at length in our guide to evaluating agentic AI vendors. A system operating under a well-drawn mandate is a colleague. A system operating without one is a liability that has not yet found its occasion.

The mandate questions are unglamorous. Which steps does the system complete on its own, and which does it prepare for a person. What does it do when it encounters something outside its brief. What can it never do, regardless of how it is configured. Who can change those limits. None of this is technically interesting, and all of it determines whether the deployment is safe.

A word about AGI

You will encounter the term. Artificial general intelligence refers to a hypothetical system with broad, human-level capability across arbitrary problems, as opposed to the narrow systems that exist today.

It has not been achieved, the term has no agreed definition, and it is best treated as a research goal rather than a product category. If a vendor invokes AGI in a commercial conversation, they are selling you something other than what they are describing. The systems available to your business are narrow, capable, and specific, and that is quite enough to be getting on with.

What we are still working out

We do not know how much autonomy enterprises will ultimately accept. Our bet, and it is visible in how we build, is considerably less than the market currently assumes. We think the durable pattern is a system that does the great majority of the work and stops at the point where judgement carries consequence, with a person in the path at that point. It is possible we are being conservative and that trust will build faster than we expect. If so, our architectures will need to loosen.

We are also uncertain about the long-run shape of multi-agent systems. Coordinating several agents to pursue a goal works, and it works better than a single agent for complex tasks. It also multiplies the places where a mandate can be breached, and the governance thinking on this is well behind the engineering. We are working in this territory, and we would rather say plainly that the answers are not settled than pretend our approach is finished.

—

Mitochondria is an agentic AI product company based in Amsterdam and Pune. ISO 27001:2022 certified. Classified within the limited and minimal risk tiers of the EU AI Act, with controls aligned to the GDPR, UK GDPR and India's DPDP Act.